What's New in 7.0.RC2

Operator Development Tools

• The msProf tool supports multi-operator collection and memory heatmap.
• The memory and contention detection capabilities of the msSanitizer tool are enhanced.

Analysis and Migration

The PyTorch GPU2Ascend tool supports PyTorch 2.2.0.

msModelSlim

• The m4 and m5 algorithms are added for outlier suppression in foundation model quantization.
• SD XL quantization is supported.
• The W8A8 tuning mode is added for foundation model quantization.
• The per_token, kv-cache, W4A16, and W4A16 per_group quantization capabilities for foundation model quantization are added.

Performance Tuning Tool

• Memory information can be collected and parsed at the operator granularity in MindSpore scenarios..
• Dynamic profiling is supported in the PyTorch scenario.
• Customized dotting is supported in the PyTorch scenario.
• Profiling supports operator attribute collection.

MindStudio Insight

• Linux (glibc >= 2.27) is supported.
• MindSpore memory data can be displayed.
• Automatic analysis of profile data is supported and modification suggestions are provided.
• Functions such as operator memory load and computing load display are added for operator-level tuning.

IDE User Guide

The TensorFlow GPU2Ascend function is deleted.